Re: [PATCH v2 5/7] virtio-mem: s390 support

[David Hildenbrand <david@redhat.com>]

Am 21.10.24 um 08:33 schrieb Christian Borntraeger:
> 
> 
> Am 15.10.24 um 10:37 schrieb Heiko Carstens:
>> On Mon, Oct 14, 2024 at 09:16:45PM +0200, David Hildenbrand wrote:
>>> On 14.10.24 20:48, Heiko Carstens wrote:
>>>
>>> The cover letter is clearer on that: "One remaining work item is kdump
>>> support for virtio-mem memory. This will be sent out separately once initial
>>> support landed."
>>>
>>> I had a prototype, but need to spend some time to clean it up -- or find
>>> someone to hand it over to clean it up.
>>>
>>> I have to chose wisely what I work on nowadays, and cannot spend that time
>>> if the basic support won't get ACKed.
>>>
>>>
>>> For many production use cases it certainly needs to exist.
>>>
>>> But note that virtio-mem can be used with ZONE_MOVABLE, in which case mostly
>>> only user data (e.g., pagecache,anon) ends up on hotplugged memory, that
>>> would get excluded from makedumpfile in the default configs either way.
>>>
>>> It's not uncommon to let kdump support be added later (e.g., AMD SNP
>>> variants).
>>
>> I'll leave it up to kvm folks to decide if we need kdump support from
>> the beginning or if we are good with the current implementation.
> 
> If David confirms that he has a plan for this, I am fine with a staged approach
> for upstream.

I do have a plan and a even a semi-working prototype that I am currently 
improving. In summary, the virtio-mem driver in kdump mode can report ranges 
with plugged memory to the core so we can include them in the elfcore hdr. That 
is the easy part.

The "challenge" is when the virtio-mem driver is built as a module and gets 
loaded after building/allocating the elfcore hdr (which happens when creating 
/proc/vmcore). We have to defer detecting+adding the ranges to the time 
/proc/vmcore gets opened. Not super complicated, but needs some thought to get 
it done in a clean way / with minimal churn.

-- 
Cheers,

David / dhildenb